The Astrophysical Journal, 732:82 (11pp), 2011 May 10 doi:10.

1088/0004-637X/732/2/82

C 2011. The American Astronomical Society. All rights reserved. Printed in the U.S.A.

CONSTRAINING THE DARK ENERGY EQUATION OF STATE USING LISA OBSERVATIONS OF SPINNING
MASSIVE BLACK HOLE BINARIES
Antoine Petiteau, Stanislav Babak, and Alberto Sesana
Max-Planck-Institut fuer Gravitationsphysik, Albert-Einstein-Institut, Am Muhlenberg 1, D-14476 Golm bei Potsdam, Germany
Received 2010 December 10; accepted 2011 March 1; published 2011 April 20

ABSTRACT
Gravitational wave (GW) signals from coalescing massive black hole (MBH) binaries could be used as standard
sirens to measure cosmological parameters. The future space-based GW observatory Laser Interferometer Space
Antenna (LISA) will detect up to a hundred of those events, providing very accurate measurements of their luminosity
distances. To constrain the cosmological parameters, we also need to measure the redshift of the galaxy (or cluster of
galaxies) hosting the merger. This requires the identification of a distinctive electromagnetic event associated with
the binary coalescence. However, putative electromagnetic signatures may be too weak to be observed. Instead, we
study here the possibility of constraining the cosmological parameters by enforcing statistical consistency between
all the possible hosts detected within the measurement error box of a few dozen of low-redshift (z < 3) events.
We construct MBH populations using merger tree realizations of the dark matter hierarchy in a ΛCDM universe,
and we use data from the Millennium simulation to model the galaxy distribution in the LISA error box. We show
that, assuming that all the other cosmological parameters are known, the parameter w describing the dark energy
equation of state can be constrained to a 4%–8% level (2σ error), competitive with current uncertainties obtained
by type Ia supernovae measurements, providing an independent test of our cosmological model.
Key words: black hole physics – cosmological parameters – galaxies: distances and redshifts – gravitational waves
– methods: statistical
Online-only material: color figures

1. INTRODUCTION in order to measure the source redshift and be able to do

The Laser Interferometer Space Antenna (LISA; Danzmann &
the LISA Study Team 1997) is a space-based gravitational wave
(GW) observatory which is expected to be launched in 2022+.
One of its central scientific goals is to provide information
about the cosmic evolution of massive black holes (MBHs). It
is, in fact, now widely recognized that MBHs are fundamental
building blocks in the process of galaxy formation and evolution;
they are ubiquitous in nearby galaxy nuclei (see, e.g., Magorrian
et al. 1998), and their masses tightly correlate with the properties
of their host (Gültekin et al. 2009, and references therein). In
popular ΛCDM cosmologies, structure formation proceeds in a
hierarchical fashion (White & Rees 1978), through a sequence
of merging events. If MBHs are common in galaxy centers at
all epochs, as implied by the notion that galaxies harbor active
nuclei for a short period of their lifetime (Haehnelt & Rees
1993), then a large number of MBH binaries are expected to
form during cosmic history. LISA is expected to observe the
GW-driven inspiral and final coalescence of such MBH binaries
out to very high redshift with a high signal-to-noise ratio (S/N),
allowing very accurate measurements of the binary parameters.
The collective properties of the set of the observed coalescing
binaries will carry invaluable information for astrophysics,
making it possible to constrain models of MBH formation and
growth (Plowman et al. 2010; Gair et al. 2010; Sesana et al.
2010).
Besides astrophysical applications, coalescing MBHs could
be used as standard sirens (Schutz 1986; Holz & Hughes 2005;
Lang & Hughes 2006, 2009; Arun et al. 2007, 2009a; Van Den
Broeck et al. 2010). The high strength of the GW signals allows
us to measure the luminosity distance with a precision of less
than a percent at redshift z = 1 (neglecting weak lensing).
However, we need an electromagnetic identification of the host
cosmography. If the event is nearby (z < 0.4), then we have
a very good localization of the source on the sky and we can
identify a single cluster of galaxies hosting the merger. As we
go to higher redshifts, LISA sky localization abilities become
quite poor: a typical sky resolution for an equal mass 106 M
inspiralling MBH binary at z = 1 is 20–50 arcmin per side at
2σ (Trias & Sintes 2008; Lang & Hughes 2009; Arun et al.
2009b), which is in general not sufficient to uniquely identify
the host of the GW event. There is, therefore, a growing interest
in identifying putative electromagnetic signatures associated
with the MBH binary before and/or after the final GW-driven
coalescence (for a review, see Schnittman 2010, and references
therein). Electromagnetic anomalies observed before or after
the coalescence within the LISA measurements error box may
allow us to identify the host and to make a redshift measurement.
However, most of the proposed electromagnetic counterparts
are rather weak (below the Eddington limit), and in case of
dry mergers (no cold gas efficiently funneled into the remnant
nucleus) we do not expect any distinctive electromagnetic
transient. This brings us back to the original idea by Schutz
(1986) of considering each galaxy within the LISA measurement
error box as a potential host candidate. The idea is that, by
cross-correlating several GW events, only one galaxy (cluster
of galaxies) in each error box will give us a consistent set of
parameters describing the universe. The effectiveness of this
method has been demonstrated by MacLeod & Hogan (2008) in
the context of the Hubble constant determination by means of
low-redshift (z < 0.2) extreme mass ratio inspirals.
We use the hierarchical MBH formation model suggested by
Volonteri & Begelman (2010) to generate catalogs of coalescing
MBH binaries along the cosmic history. This model predicts
∼100 MBHs mergers observable by LISA in three years, in
the redshift range [0:5]. We do not use sources beyond redshift

1
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
z = 3 due to difficulties of measuring galaxy redshifts beyond
that threshold.1 We model the galaxy distribution in the universe
using the Millennium simulation (Springel et al. 2005). For each
coalescing MBH in our catalog, we select a host galaxy in the
Millennium run snapshot closest in redshift to the actual redshift
of the event. For each galaxy in the snapshot, we compute
the apparent magnitude in some observable band, and we
create a catalog of redshift measurements of all the observable
potential host candidates. Note that typical observed mergers
involve 104 –106 M MBHs, which implies (using the black hole
mass–bulge relations; see, e.g., Gültekin et al. 2009) relatively
light galaxies. However, observed galaxies are heavy due to
selection effects: roughly speaking, mass reflects luminosity,
so that at high redshifts we can observe only very massive
(luminous) galaxies. Therefore, the actual host might not be (and
often is not) among the observed galaxies. The important fact is
the self-similarity of the density distribution: the local density
distribution for all galaxies and the density distribution for heavy
galaxies are quite similar, which allow us to infer the likelihood
of the host redshift on the basis of redshift measurements of the
luminous galaxies only.
We assume that the GW source parameter measurements (GW
likelihoods) are represented by multivariate Gaussian distribu-
tions around the true values, with the variance–covariance ma-
trix defined by the inverse of the Fisher matrix. This is a good
approximation in the case of Gaussian instrumental noise and
large S/N. At z  0.25, the uncertainty in the luminosity dis-
tance (DL ) is dominated by weak lensing (WL) due to the ex-
tended distribution of dark matter (DM) halos between us and
the GW source. In this paper, we combine the luminosity dis-
tance errors given by GW measurements and WL, referring to
them as GW+WL errors. We use two estimations of the WL
error (1) from Shapiro et al. (2010) and (2) from Wang et al.
(2002).
In order to evaluate the error box, we need to assume some
prior on the cosmological parameters. In this exploratory study,
we assume that we know all the cosmological parameters but
the effective equation of state for the dark energy, described
by the parameter w (which could be the case by the time LISA
will fly). In a follow-up paper, we will relax this assumption
by including also the Hubble constant and the matter and dark
energy content of the universe as free parameters. We take the
prior range for w from the seven year Wilkinson Microwave
Anisotropy Probe (WMAP) analysis (Komatsu et al. 2011). We
show that using statistical methods w can be constrained to
a 4%–8% level (2σ error), providing an effective method for
estimating the dark energy equation of state. We also show that
this result depends weakly on the prior range and could serve as
an independent way of measuring the dark energy equation of
state, with respect to canonical methods employing observations
of type Ia supernovae (SNe Ia; Riess et al. 1998).
The paper is structured as follows. In Section 2, we explicitly
spell out all the details of the adopted cosmological model
and of the Bayesian analytical framework. In Section 3, we
give more insights on the MBH population model and on the
galaxy distributions extracted from the Millennium database. In
Section 4, we describe our simulated GW and electromagnetic
observations. We give results of our simulations under different
assumptions about WL, depth of the follow-up electromagnetic
surveys, etc., in Section 5. We summarize our findings in
Section 6.
1 There are other reasons for not going beyond z = 3 which we will discuss
later.
2
Petiteau, Babak, & Sesana
2. ANALYTICAL FRAMEWORK
2.1. Cosmological Description of the Universe
We assume the standard ΛCDM model, which describes
our universe as the sum of two non-interacting components:
(1) a pressureless component corresponding to all visible and
DM, and (2) a dark energy component with current effective
equation of state ω(z) (for the standard Λ-term, p = −,
it would correspond to ω = −1). Current estimates based
on SN Ia observations and anisotropy measurements in the
cosmic microwave background (Riess et al. 1998; Komatsu
et al. 2011) tell us that about 70% of the universe en-
ergy content is in the form of the dark energy. The evolu-
tion of the universe is therefore described by the expansion
equation
   z 
1 + ω(z)
H = H0 Ωm (1 + z) + Ωde exp 3
2 2 0 3 0
dz , (1)
0 1+z
where H = ȧ/a (a being the lengthscale of the universe) is the
Hubble expansion parameter and H0 is its current value (t = 0),
Ωm and Ωde are the ratios of the matter density and the dark
energy density to the critical density, and ω(z) describes the
effective dark energy equation of state as a function of z. We
assume that the universe is spatially flat, the luminosity distance
is therefore computed as
 z
dz
DL = (1 + z) . (2)
0 H (z )
In our simulations, we fix all parameters (assuming that they
are known exactly) to the currently estimated mean values:
H0 = 73.0 km s−1 Mpc−1 , Ωm = 0.25, and Ωde = 0.75.
We also simplify the form of ω(z) for which we will assume
ω = −1 − w, where w is a constant.2 We choose the value
w = 0 to simulate our universe which is what has been used in
the Millennium simulation (see below).
2.2. Methodology and Working Plan
Our aim is to show that we can constrain w via GW obser-
vations of spinning MBH binaries using a Bayesian framework.
Let us consider j = 1, . . . , Nev GW observations. For each
event, we can infer the probability of a parameter w, given the
collected data s, using Bayes theorem:
p0 (w)Pj (s|w)
Pj (w|s) = . (3)
Ej
Here, Pj (w|s) is the posterior probability of the parameter
w, Pj (s|w) is the likelihood of the parameter w given the
observation s, p0 (w) in the prior knowledge of w and Ej is
defined as 
Ej = p0 (w)Pj (s|w)dw. (4)
The likelihood Pj (w|s) must be appropriately specialized to
our problem. We want to exploit GW observations to constrain
w through the distance–redshift (DL –z) relation as given by
Equation (2).
2 Here, we use notations for the dark energy equation of state adopted in the
WMAP data analysis (Komatsu et al. 2011).
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
1. The distance DL is provided by the GW observations: the
GW signal carries information about the parameters of the
binary, including its location on the sky and its luminosity
distance. All of those parameters can be extracted using
latest data analysis methods (Petiteau et al. 2010; Cornish
& Porter 2007). The measurements errors are encoded in
the GW likelihood3 function L(DL , θ, φ, λ), where {θ, φ}
are the ecliptic coordinates of the source and λ  represents
all of the other parameters characterizing the MBH binary
(spin and orientation, mass, orientation of the orbit, and
the MBH’s position at the beginning of observations).
When estimating DL , WL cannot be neglected. In fact,
the error coming from the WL (causing fluctuations in the
brightness of the GW source, which gives an uncertainty
in the luminosity distance) dominates over the GW error
starting from redshift z ∼ 0.25 (see Figure 2).
2. The redshift measurement does not rely on any distinctive
electromagnetic signature related to the GW event. We
extract a redshift probability distribution of the host from
the clustering properties of the galaxies falling within the
GW+WL error box. This defines an astrophysical prior
p(θ, φ, z) for a given galaxy in the measurement error
box to be the host of coalescing binary. To translate the
measured DL and uncertainty ΔDL of the GW event into
a corresponding z and Δz for the candidate host galaxies
in the sky, we use the prior knowledge of p0 (w) obtained
from WMAP.
The likelihood in Equation (3) can therefore be written as
 through a sequence of merger and accretion episodes. Two dis-
 ]p(λ
 )pj (θ, φ, z)
Pj (s|w) = Lj [DL (z, w), θ, φ, λ
 dθ dφ dz,
× dλ
 on the parameters λ 
where we have introduced the priors p(λ)
(which we assume in this paper to be uniform). It is convenient
to change the variable of integration from DL to z. Since we have
assumed uniform priors on λ,  we can marginalize the likelihood
over those parameters4 to obtain
 Begelman (Volonteri & Begelman 2010, hereafter VB model),
Pj (s|w) = πj [DL (z, w), θ, φ] pj (θ, φ, z) dθ dφ dz, (6)
where we denoted the marginalized GW likelihood as
πj [DL (z, w), θ, φ]. Practically, we limit the integration to the
size of the error box (in principle the integration should be taken
over the whole range of parameters but we found that consider-
ing the 2σ error box is sufficient).
We assume that the error in luminosity distance from the WL
is not correlated with the GW measurements, hence the integral
in Equation (6) can be performed over the sky ({θ, φ}) first, and
then over the redshift. We also found that the correlation between
DL and the sky position coming from the GW observations is
not important for events at z < 0.5. Plugging Equation (6)
into Equation (3) defines the posterior distribution of w for a
single GW event (as indicated by the index j). Assuming that
3 ing that these models successfully reproduce several properties
Through the paper, with GW likelihood we mean the likelihood of the LISA
data to contain the GW signal with a given parameters, not to be confused with
the likelihood Pj (s|w) defined in the Bayes theorem.
4 Here, this corresponds to the projection of the Fisher matrix to
three-dimensional parameter space of sky location θ, φ, and luminosity
distance DL .
Petiteau, Babak, & Sesana
all Nev GW events are independent, the combined posterior
probability is
 ev
p0 (w) Nj =1 Pj (s|w)
P (w) =  Nev . (7)
p0 (w) j =1 Pj (s|w)dw
To evaluate w through Equation (7) we therefore need the
following.
1. An MBH binary population model defining the properties
of the Nev coalescing systems;
2. The spatial distribution of galaxies within a volume com-
parable with the combined GW+WL measurement error
box;
3. The measurement errors associated with GW observations

of coalescing MBH binaries (defining Lj (DL , θ, φ, λ));
4. An estimation of spectroscopic survey capabilities to con-
struct the galaxy redshift distribution within the GW+WL
measurement error box (defining pj (θ, φ, z)).
We will consider these points individually in the next two
sections.
3. ASTROPHYSICAL BACKGROUND
3.1. Massive Black Hole Binary Population
To generate populations of MBH binaries in the universe,
we use the results of merger tree simulations described in de-
tails in Volonteri et al. (2003). MBHs grow hierarchically, start-
ing from a distribution of seed black holes at high redshift,
tinctive type of seeds have been proposed in the literature. Light
(M ∼ 100 M ) seeds are thought to be the remnant of Popu-
lation III (POPIII) stars (Madau & Rees 2001), whereas heavy
(5)
seeds form following instabilities occurring in massive proto-
galactic disks. In the model proposed by Begelman Volonteri
& Rees (Begelman et al. 2006, hereafter BVR model), a
“quasistar” forms at the center of the protogalaxy, eventually
collapsing into a seed BH that efficiently accretes from the qua-
sistar envelope, resulting in a final mass M ∼ few ×104 M .
Here, we use the model recently suggested by Volonteri &
which combines the two above prescriptions by mixing light
and heavy initial seeds. This model predicts ∼30–50 events per
year in the redshift range 0 < z < 3, relevant to this study.
The dashed blue lines in Figure 1 show the redshifted total mass
(Mz = (M1 + M2 )(1 + z), being M1 > M2 the rest-frame masses
of the two MBHs, upper left panel), mass ratio (q = M2 /M1 ,
upper right panel), and redshift (lower panel) distribution of
the MBH binaries coalescing in three years, as seen from the
Earth. The model predicts ∼40 coalescences in the redshifted
mass range 105 M < Mz < 107 M , almost uniformly dis-
tributed in the mass ratio range 0.1 < q < 1, with a long
tail extending to q < 10−3 . For comparison, we also show the
population expected by a model featuring a heavy seed only
(BVR model, green dot-dashed lines), and by an alternative
VB type model (labeled VB-opt for optimistic, red long-dashed
lines) with a boosted efficiency of heavy seed formation (see
Volonteri & Begelman 2010 for details). It is worth mention-
of the observed universe, such as the present-day mass density
of nuclear MBHs and the optical and X-ray luminosity func-
tions of quasars (Malbon et al. 2007; Salvaterra et al. 2007).
The BVR and the VB-opt models predict MBH population ob-
servables bracketing the current range of allowed values. The
3
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
Figure 1. Population of coalescing MBH binaries in three years. Top left panel:
total redshifted mass distribution; top right panel: mass ratio distribution; lower
panel: redshift distribution. Color and linestyle codes are labeled in the figure.
(A color version of this figure is available in the online journal.)
VB-opt model, in particular, is borderline with current observa-
tional constraints on the unresolved X-ray background, and it is
shown here only for comparison. In the following, we consid-
ered the VB model only, which fits all the relevant observables
by standing on the conservative side.
We performed 100 Monte Carlo realizations of the population
of MBH binaries coalescing in three years. Each realization
takes into account the distribution of the number of events and
MBH masses with the redshift as predicted by the VB model.
Other parameters (like time of coalescence, spins, initial orbital
configuration) are chosen randomly using uniform priors over
the appropriate allowed ranges.
3.2. Galaxy Distribution
To simulate the galaxy distribution in the universe, we use
the data produced by the Virgo Consortium publicly available
at http://www.g-vo.org/Millennium. These data are the result of
the implementation of semianalytic models for galaxy formation
and evolution into the DM halo merger hierarchy generated
by the Millennium simulation (Springel et al. 2005). The
Millennium run is an N-body simulation of the growth of DM
structures in the expanding universe starting from a Gaussian
spectrum of initial perturbations in the DM field at high redshift,
which successfully reproduced the net-like structure currently
observed in the local universe. The simulation has a side
length of ≈700 Mpc (comoving distance), and its outcome is
stored in 63 snapshots evenly separated in log(z), enclosing
all the properties of the DM structure at that particular time.
Semianalytical models for galaxy formation are implemented a
posteriori within the DM structures predicted by the simulation.
Such models have been successful in reproducing several
observed properties of the local and the high-redshift universe
(see, e.g., Bower et al. 2006; De Lucia & Blaizot 2007). Here, we
use the implementation performed by Bertone and collaborators
(Bertone et al. 2007), which is a refinement of the original
implementation by De Lucia & Blaizot (2007).
For each coalescing MBH binary, we choose the snapshot
closest in redshift. Within the snapshot we choose the host of
4
Petiteau, Babak, & Sesana
the GW signal according to a probability proportional to the
number density of neighbor galaxies ngal . Such an assumption
comes from the fact that two galaxies are needed in order to
form an MBH binary, and we consider that the probability that a
certain galaxy was involved in a galaxy merger is proportional to
the number of neighbor galaxies. We consider to be neighbors of
a specific galaxy all the N (R) galaxies falling within a distance
R = σ TH (z), (8)
where σ = 500 km s−1 is the typical velocity dispersion
of galaxies with respect to the expanding Hubble flow, and
TH (z) is the Hubble time at the event redshift. The number
density of neighbor galaxies is then simply written as ngal =
3N (R)/(4π R 3 ). When we choose the merger host, we compute
ngal considering all the neighbor galaxies, without imposing any
kind of mass or luminosity selection. In this case ngal ≡ ntotal .
However, when we will construct the probability of a given
observable galaxy to be the host of the merger (i.e., the
astrophysical prior pj (θ, φ, z)), we will have to compute ngal
according to the number of observed neighbors, because this
is the only thing we can do in practice when we deal with an
observed sample of galaxies (see Section 4.2).
4. SIMULATING THE OBSERVATIONS
4.1. Gravitational Wave Observations: Shaping the Error Box
As we mentioned in Section 3.1, we performed 100 realiza-
tions of the MBH binary population from the VB model. Each
realization contains 30–50 events in the redshift range [0:3]. The
total mass, mass ratio, and redshift distributions of the events
are shown in Figure 1. In order to simulate GW observations,
the binary sky location is randomly chosen according to a uni-
form distribution on the celestial sphere, the coalescence time
is chosen randomly within the three years of LISA operation
(we assume three years as default mission lifetime). The spin
magnitudes are uniformly chosen in the interval [0:1] in units
of mass squared, and the initial orientations of the spins and of
the orbital angular momentum are chosen to be uniform on the
sphere. More detailed description of the model for GW signal
used in this paper is given in Petiteau et al. (2010).
The GW likelihood L needed in Equation (5) is approximated
as a multivariate Gaussian distribution with inverse correlation
matrix given by the Fisher information matrix (FIM):
L ∼ e−(s−h|s−h) ∼ e(θ −θ̂ i )Γij (θ j −θ̂ j )/2
i
. (9)
Here, θ i is the vector of the parameters characterizing the
spinning MBH binary, θ̂ i are the maximum likelihood estimators
for those parameters which are assumed to correspond to
the true values (no bias), and Γij = (h,i |h,j ) is the FIM,
where the commas correspond to derivatives with respect to
the parameters. This is a reasonable approximation due to the
large S/N (for more details on the FIM and its applicability,
see Vallisneri 2008). Our uncertainties on estimated parameters
are consistent with Lang & Hughes (2009), Babak et al. (2010),
and Petiteau et al. (2010). We did not include higher harmonics
(only the dominant, twice the orbital frequency) as they only
slightly improve parameter estimation for precessing binaries.
However, including higher harmonics in the GW signal model is
important in the case of the small spins and low precession (when
spins are almost (anti)aligned with the orbital momentum; Lang
et al. 2011). We use truncated waveforms corresponding to the
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
0.06
0.05
GW error
Weak lensing error #1
Weak lensing error #2
Combined error #1
0.04 Combined error #2
δDL/DL
0.03
0.02
0.01
0 100
0 0.5 1 1.5 2 2.5
z 150
Figure 2. Relative error in the luminosity distance due to WL from
(1) Shapiro et al. (2010, circles) and from (2) Wang et al. (2002, squares). The
black solid line is the median error due to GW measurements only; the solid-
circle and the solid-square lines are for the combined errors under assumptions
(1) and (2), respectively (see the text).
(A color version of this figure is available in the online journal.)
inspiral only. However, the addition of a merger and ringdown
will further reduce the localization error due to the higher S/N
(McWilliams et al. 2010). This error is usually an ellipse on
the sky but we simplify it by choosing the circle with the same
area.
For the luminosity distance measurement, we need to take into
account the WL. We assume the WL error to be Gaussian with a
σ given by (1) Shapiro et al. (2010). Such assumption is rather
pessimistic; we also tried the prescription given by (2) Wang
et al. (2002), which gives smaller errors, but still larger than the
level that may be achieved after mitigation through shear and
flexion maps (Hilbert et al. 2011). Both of those estimations are
represented in Figure 2 as (1) dark (red online) circles and (2)
light (orange online) squares correspondingly. The median error
in DL due to GW measurements only is given by the solid black
line. The combined error for model (1) is given by the upper
(blue) circle-line curve, and for model (2) by the lower (green)
square-line curve. We consider our setup to be conservative in
the estimation of the WL effects. The main aim of this work
is to build a reasonable setup for what could be observed by
the time LISA will fly, and make a first order estimation of
LISA capabilities to constrain the dark energy equation of state.
We will address non-Gaussianity of the WL as well as other
corrections to the model to make it more realistic in a follow-up
paper.
We consider an error box size corresponding to 2σ of the
measurement errors in the sky location (σsky ) and in the source
distance as evaluated by the FIM plus WL uncertainties. For
observational purposes, the dimensions of this error box are
ΔΩ = 2σsky and Δz. For the latter, we also include the
uncertainty given in the Dl –z conversion due to the error (prior)
on w, p0 (w).
Let us summarize how we construct an error box in practice,
as, for example, the one illustrated in Figure 3.
1. We select the closest Millennium snapshot to the event in
redshift.
Petiteau, Babak, & Sesana
400
350
300
250
400
200
150 350
100
50 300
250
50 200
3
150
200
250 100
300
350
400 50
Figure 3. Example of error box (cylinder) in part of the Millennium snapshot
(cube with unit in Mpc). The blue cylinder is the measurement error box and
the green one also considers the prior on w. The black big dot is the host and
the brown small dots are the selected galaxy candidates.
(A color version of this figure is available in the online journal.)
2. We pick a galaxy (red dot) in the snapshot with a probability
given by the local galaxy number density ntotal .
3. We construct around the galaxy an error box given by ΔΩ
and ΔzGW+WL , and the galaxy can lie anywhere with respect
to this error box (blue cylinder).
4. We expand the error box along the direction of the observer
both sides by Δz given by the uncertainty in w (green
cylinder).
5. According to some prescription, which we will describe in
the next section, we select observable galaxies in the error
box (brown dots).
As shown in Figure 3, we interpret one of the directions
in the Millennium snapshot as distance from the observer, and
convert the comoving distance in redshift. We assume a periodic
expansion of the Millennium data in order to fit large error boxes.
Note that the original Millennium simulation also assumes the
same periodicity in the distribution of the matter. The size of
the error box at high redshift covers a significant fraction of
the simulation box so we do not go beyond the redshift z = 3
(as we will show later, spectroscopic observations at such high
redshifts will be impractical anyway). Together with larger error
boxes, we have a nonlinear increase in the number of events
at high redshift. To reduce the overlap between error boxes
corresponding to different GW events, we choose cylinders with
random orientations.
Figure 4 shows an example of the resulting weighted distri-
bution of galaxy redshifts (with weight proportional to the local
density ntotal ). It is a projection of the clumpiness along the line
of sight which is also proportional to the probability distribution
of z for the event. The probability distribution of w for the event
will be directly related to this result. We noticed that there is a
very large number of underdense regions and several very dense
superclusters. The probability of a galaxy with a low local den-
sity to host a merger is very low but there is a huge number of
such galaxies, and we found that the probability of the host to
be in (super)clusters is similar to that of being in a low-density
5
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
3 tering information we need by getting spectra of the brightest
2.5
2 semianalytic galaxy evolution model (Bertone et al. 2007)
scaled local density
1.5
1
0.5
0
1.2 1.25 1.3 1.35 1.4 1.45
z
Figure 4. Distribution of the weighted galaxies with the redshift. The green
dashed vertical line is the redshift of the host galaxy.
(A color version of this figure is available in the online journal.)
region. As we will see later in the result section, this may cause
a very wrong estimation of w for some individual GW event.
4.2. Redshift Measurements Through Spectroscopic Surveys
To get a statistical measurement of w, we need to exploit the
clustering of the galaxies falling within the error box (which
defines the astrophysical prior pj (θ, φ, z) in Equation (5)). It
is therefore necessary to get efficient redshift measurements of
thousands of galaxies within a small field of view (FOV): the in-
formation we seek is enclosed in the redshift distribution of such
galaxies. We stress here that we are not looking for a distinctive
electromagnetic counterpart to the GW event. In fact, the actual
host of the coalescing binary may not even be observable. Typi-
cal masses of our binaries are 105 –106 M . Using MBH–bulge
scaling relations (Gültekin et al. 2009), such MBHs are expected
to be hosted in galaxies with stellar mass 109 –1010 M , i.e., in
DM halos with total mass <1011 M . The Millennium run mass
resolution is ∼109 M , meaning that typical host structures are
formed by less than 100 particles. Unfortunately, the Millen-
nium run is severely incomplete in the expected mass range of
LISA MBH binary hosts. Here, we do not attempt to exploit any
MBH–host relation to select the host of our GW event; the prob-
ability of being a host is only related to the local number density
of neighbor galaxies ntotal . Such an assumption relies on the con-
cept of self-similarity of the galaxy clustering at different mass
scales: typical LISA MBH binary hosts cluster in the same way as
more massive galaxies. We checked this assumption by compar-
ing the spatial distribution of galaxies in different mass ranges
(109 –1010 M , 1010 –1011 M , 1011 –1012 M ), within simula-
tion snapshots at different redshift, and we postulate that this
self-similarity extends to lower masses, below the Millennium
run resolution. This point is crucial for two reasons: (1) espe-
cially at z > 1, we will be able to get only spectra of luminous
(massive) galaxies, and we need to be confident that their spatial
distribution mimics that of lighter galaxies that may host the GW
event but are observable in the spectroscopic survey and (2) the
number of observable galaxies in the error box may be too large
anyway (>104 ) to efficiently complete a spectroscopic survey
on the full sample: self-similarity allows us to get the clus-
Petiteau, Babak, & Sesana
objects only.
At z = 1, the typical number of galaxies enclosed in the 2σ
error box described above is in the range 104 –105 . However,
not all of them are bright enough to get useful spectra. The
implemented on top of the Millennium run returns the stellar
mass of each galaxy, and the absolute bolometric magnitude
Mb . By knowing the redshift, and by using standard galactic
templates one can therefore compute the apparent magnitude in
a given band, by assuming the appropriate k-correction (Oke &
Sandage 1968). Here, we use the R-band apparent magnitude
mr for illustrative purposes, and we adopt the relation (Zombeck
1990)
Mb = −5log(zc/H0 ) − 1.086z − 25 + mr + 0.6, (10)
1.5 where 0.6 is the k-correction. For each galaxy, we compute mr
and we simulate spectroscopic surveys at different thresholds
mr = 24, 25, 26. We stress here that the GW host was
chosen among all of the galaxies falling in the error box, and
therefore may not (and usually does not) belong to the observed
sample. We then assume that for each galaxy satisfying the
survey threshold we get an exact spectroscopic redshift, and we
combine the redshift distribution of several error boxes to get a
statistical estimation of w. In practice, each redshift estimation
will come with a measurement error, and an intrinsic error due
to the proper motion of the source with respect to the Hubble
flow. Both errors are, however, of the order of Δz/z < 10−3 ,
well below the typical redshift scale corresponding to spatial
clustering of structures (Δz ∼ 0.01; see Figure 4) we need to
resolve.
Our method does not rely on the observation of a prompt
transient associated with the MBH binary coalescence to iden-
tify the host galaxy. Nevertheless, getting thousands (or tens
of thousands) of spectra in a small FOV requires a dedicated
observational program. Thanks to multi-slit spectrographs such
as VIMOS at Very Large Telescope (VLT; Le Fèvre et al. 2003)
and DEIMOS at Keck (Faber et al. 2003), fast deep spectro-
scopic surveys of relatively large FOV are now possible. For
example, the ongoing VIMOS VLT deep survey (Le Fèvre et al.
2005), took spectra of >10,000 galaxies, mostly in the redshift
range 0 < z < 1.5, within an FOV of 0.61 deg2 at an apparent
magnitude limit IAB < 24. Comparable figures are achieved by
other observational campaigns such as zCOSMOS (Lilly et al.
2009) and DEEP2 (Davis et al. 2003), which were able to survey
selected galaxies in various photometric bands (U, B, R, I ) to
an apparent magnitude limit of about 24. Going deeper in red-
shift, Lyman break galaxy redshift surveys are finding success in
efficiently getting high-quality spectra of hundreds of galaxies
in the redshift range 2.5 < z < 3.5 within an FOV ∼ 1 deg2
(Bielby et al. 2010). To illustrate this, the VIMOS spectrograph
can take ∼500 high-quality spectra per pointing with an inte-
gration time of about 4 hr, within a 7 × 8 arcmin2 FOV, which
is coincidentally of the same order of the typical error box for
a z = 1 GW event. The typical redshift accuracy of the spectra
is Δz < 10−3 (3 × 10−4 in the zCOSMOS survey, 2 × 10−3 in
the Lyman break galaxy survey), well below the typical redshift
scale we are interested in (z ∼ 0.01).
Such results are a testament to the feasibility of efficient
spectroscopic redshift determination of a large sample of galax-
ies at faint apparent magnitude (mr ≈ 24), as required by our
study. The future spectroscopic survey BigBOSS (Schlegel et al.
2009) is expected to further improve such figures of merit; a new
6
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
spectrograph will be able to simultaneously get up to 4000 spec-
tra within a single pointing of a 7 deg2 FOV. Getting a few thou-
sand spectra of objects falling within the GW error box in the
redshift range of interest may be possible in a single observing
night. At a mr = 24 cutoff magnitude we generally have few
hundred to few thousands galaxies in the GW error box, but
we go deeper (i.e., mr = 26, feasible with future surveys), the
number of spectra may increase drastically. For some of the
error boxes, we count up to 105 galaxies with mr < 26. How-
ever, the requirement of a factor of 10 more spectra does not
correspond to a significant improvement of the results. This is a
consequence of the self-similarity of the galaxy distribution: as
long as there are enough galaxies in the error box to recover the
clustering information, the results are basically independent on
the assumed cutoff magnitude. A survey with a cutoff magnitude
of mr = 24 may indeed be a good compromise between relia-
bility of the results and time optimization in terms of follow-up
spectroscopy.
The magnitude cutoff defines the number of neighbor ob-
servable galaxies. This is the only practical way to weight each
galaxy with a local density, ngal ≡ nmr (the subscript mr refers
to the adopted magnitude limit) along the lines discussed in
Section 3.2. Once we have a spectroscopic galaxy sample, each
galaxy in the error box comes with the prior probability to
be the host proportional to nmr , so the astrophysical prior in
Equation (6) could be written as
p(Ω, z) = nmr ,i δ(Ω − Ωi )δ(z − zi ), (11)
i
where the sum is over all observable galaxies in the error box
and Ω is the geodesic distance on the celestial sphere from
the center of the box. At redshifts z  1 the prior probability
p(Ω, z) becomes almost a continuous function (as the example
in Figure 4).
4.3. Approximations and Caveats
Before discussing the results, we want to mention some
corrections we made to accommodate the limitations of our
simulations. First, we interpreted one of the directions in the
snapshot (along the side of the cylinder) as distance from the
observer. This is a good approximation only if the error box
size is small. For large error boxes, a uniform distribution
in the comoving distances does not translate into a uniform
distribution in redshifts: there is an artificial slope with a bias
toward low values of z. We have corrected for this slope. Second,
the clumpiness evolves with redshift, which is not the case if
we use a single snapshot and interpret one of the directions as a
redshift. To properly account for this, we should glue snapshots
together and perform an interpolation between them. However,
we wanted to simplify the setup for this very first attempt. The
main idea was to check whether the density contrast within the
error boxes is sufficient to further constrain the error on w.
If the distribution of density within the error box is uniform
then we do not gain any useful information. However, there
is a natural bias: for a given measurement of DL , the galaxy
further away (larger z) constrains w better than a galaxy at lower
redshift. One can see it from the fact that deviation between the
curves in DL –z plane corresponding to the small deviation in
w is larger for large z. This could be counterbalanced by the
decreasing density contrast at large redshift. Here, we corrected
the slope of the posterior Pj (w|s) by demanding that a uniform
distribution pj (θ, φ, z) returns a posterior on w equal to the
prior, i.e., Pj (w|s) = p0 (w).
7
Petiteau, Babak, & Sesana
5. CONSTRAINTS ON THE DARK ENERGY
EQUATION OF STATE
In this section, we present the results of our simulations.
We tried several setups of the experiment by using different
thresholds on the observable apparent magnitude of galaxies,
different prescriptions for the measurement errors, and different
cosmological priors. For each setup, we performed either 100
or 20 realizations of the MBH binary population as observed by
LISA, together with the follow-up spectroscopic survey of the
galaxies in all the error boxes.
5.1. Fiducial Case
We consider in this subsection 100 realizations which we refer
to as our fiducial case. For this setup, we limit spectroscopic
identification of galaxies in the error box to an apparent
magnitude of mr  24, the errors in sky localization and in the
luminosity distance are estimated according to the inspiral part
of GW signal only, and the WL uncertainty is taken from Shapiro
et al. (2010). The prior p0 (w) was assumed to be uniform
U [−0.3:0.3] with an exponential decay at the boundaries. Such
interval is consistent with current 2σ (95% confidence level)
constraints on w (w = −0.12 ± 0.27; Komatsu et al. 2011),
obtained by cross-correlating seven year WMAP data with priors
coming from independent measurements of H0 and barionic
acoustic oscillations (see Komatsu et al. 2011, and references
therein for full details), under our same assumption for the
dark energy equation of state, ω = −1 − w, where w is a
constant. Such range is reduced by a factor of almost three
(w = −0.02 ± 0.1) when SNe Ia data (Riess et al. 1998)
are included. Here, we show that GW measurements offer
a competitive alternative to SNe Ia, placing an independent
constraint on the dark energy equation of state.
We find that in almost all cases we improve the constraints
on w, in other words, the posterior distribution is narrower
than the prior. Few events at low redshift usually play a major
role in the final result. One typical realization is plotted in
the top panel of Figure 5. We split the contribution to the
posterior distribution P (w) in redshift bands: z ∈ [0:1] (second
plot from the left), [1:2] (third plot), and [2:3] (fourth plot).
Their relative contribution and the resulting posterior (black)
are given in the leftmost plot. In this example, the final posterior
probability is almost completely determined by few events
at low redshift. The second realization, shown in the lower
panels of Figure 5, demonstrates how low-redshift contributions
could give inconclusive results. In this particular case, there
are two maxima with preference given to the wrong one. The
contribution from high-redshift events could change this ratio
as it is shown in this example. In many cases, the mergers above
redshift z = 1 can constrain w only to a 0.1–0.15 accuracy, but
they almost always add up coherently, giving a maximum at the
right value (w = 0). This usually helps in case the low-redshift
events return a multimodal P (w), and is, in turn, the power of
our statistical method.
We characterize the results of each setup (100 or 20 realiza-
tions) using the figures of merit shown in Figures 6 and 7. The
first one (Figure 6) is obtained by adding the posterior distribu-
tions P (w) of all the realizations. We fit the resulting curve with
a Gaussian, characterizing the result using its mean w0 and stan-
dard deviation σw . The second figure of merit (Figure 7) shows
the result of Gaussian fits performed on each individual realiza-
tion (vertical index i): the mean w0 (i) is shown as a circle and the
standard deviation σw (i) is the error bar. The first figure of merit
The Astrophysical Journal, 732:82 (11pp), 2011 May 10 Petiteau, Babak, & Sesana

0.08 0.02 0.01

0.08
0.008
0.06 0.015
0.06
0.006
0.04 0.01
0.04
0.004

0.02 0.02 0.005

0.002

0 0 0 0
-0.3 -0.15 0 0.15 0.3
z = [0:1] z = [1:2] z = [2:3]
0.05 0.05 0.02 0.01

0.04 0.04 0.008

0.015

0.03 0.03 0.006

0.01
0.02 0.02 0.004

0.005
0.01 0.01 0.002

0 0 0 0
-0.3 -0.15 0 0.15 0.3 -0.3 -0.15 0 0.15 0.3 -0.3 -0.15 0 0.15 0.3 -0.3 -0.15 0 0.15 0.3
w w w w
Figure 5. Posterior distribution for w for two particular realizations (top and bottom row). In each row, the left plot shows the full posterior from all GW events (black
curve) as well as contributions from different redshift bands. The three right plots show the individual contribution for the three redshift ranges, as labeled in the panels.
(A color version of this figure is available in the online journal.)

(a) fiducial case (b) without electromagnetic counterpart (c) with offset gaussian prior
10 9 12
9 Data 8 Data Data
Gaussian fit : Gaussian fit : 10 Gaussian fit :
8 w0=0.0008 7 w0=0.0007 w0=0.0069
7 σ=0.0359 6 σ=0.0397 8 σ=0.0336
6
5
5 6
4
4
3 3 4
2 2
2
1 1
0 0 0
-0.3 -0.2 -0.1 0 0.1 0.2 0.3 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1
w w w
(d) apparent magnitude <26 (e) with merger (f) improved WL + merger
12 12 18
Data Data 16 Data
10 Gaussian fit : 10 Gaussian fit : Gaussian fit :
w0=0.0058 w0=-0.0067 14 w0=-0.0025
8 σ=0.0329 8 σ=0.0357 12 σ=0.0213
10
6 6
8
4 4 6
4
2 2
2
0 0 0
-0.3 -0.2 -0.1 0 0.1 0.2 0.3 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 -0.3 -0.2 -0.1 0 0.1 0.2 0.3
w w w

Figure 6. Collective figures of merit of our experiment. In each panel, corresponding to a different setup of our experiment as labeled in figure, the red solid curve
corresponds to the data, i.e., the sum of the posterior distributions of w over all realizations. The blue dashed curve is a Gaussian fit with parameter given in the legend
of each plot.
(A color version of this figure is available in the online journal.)

gives collective information, showing how well, on average, an
individual realization can be approximated by a Gaussian fit,
while the second figure of merit shows the dispersion of the
posterior distribution across the individual realizations.
The fiducial case, featuring 100 realizations, is shown in panel
(a) of both Figures 6 and 7. The parameters of the global
fitting Gaussian mean are w0 = 0.0008 and σw = 0.036,
corresponding to a factor of four improvement in the estimation
of w with respect to our standard 2σ [−0.3:0.3] prior. However,
the distribution has clearly some outliers, recognizable as non-
Gaussian tails in Figure 6 and pinned down in Figure 7. For the
fiducial case, 84% of the realizations have a mean value close
to the true one, i.e., |w0 (i) − wtrue | < 0.1 with an appreciable
reduction of the prior range, i.e., σw (i) < 0.15 (i = 1, . . . , 100

8
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
(a) fiducial case
100
80
60
index
40
20
0 0
-0.3 -0.2 -0.1 0 0.1 0.2
(e) with merger
20
15
index
10
5 5
0 0
-0.2 0 0.2
w, δw
Figure 7. Mean values and standard deviations resulting from the Gaussian fit of the posterior P (w). The setup of each panel corresponds to the one adopted in the
panel of Figure 6 labeled by the same letter).
is the realization index). Moreover, most of the outliers can be
corrected as we will explain in Section 5.6.
5.2. Removing “Electromagnetic Counterparts”
Our goal is to demonstrate that we are able to constrain
the dark energy equation of state without directly observing
electromagnetic counterparts. However, for some of the low-
redshift events, the error box is so small that only one or two
galaxies fall within it. Having one or two galaxies in the error
box essentially implies an electromagnetic identification of the
host, so we decided to re-analyze the fiducial case removing
all such fortunate events (usually 0–2 in each realization). The
fiducial case without clearly identifiable hosts is presented in
the panel (b) of Figure 6. Clearly, our results remain almost
unchanged, the posterior distribution is slightly wider (larger
sigma) and non-Gaussianity is more pronounced.
5.3. Choice of the Prior for w
Here and in the next subsections, we make use of 20 selected
realizations, which we found to be sufficient to depict the
relevant trends of the analysis. We took 15 “good” (mean values
close to the true and small rms errors) and 5 “bad” cases from
the fiducial setup.
In this subsection, we study the effect of the prior p0 (w) on
the posterior distribution. We considered an extreme case: a
Gaussian N (w0 = −0.2, σ = 0.3). As shown in panel (c) of
Figure 6, the global posterior distribution is still centered at the
true value w = 0. This demonstrates that the final conclusion
is basically unaffected by the choice of the prior (as long as the
prior covers the true value) and GW observations, in principle,
could be used as an independent mean of estimating w.
5.4. Using Deeper Surveys
Here, we study the dependence of our results on the depth
of the follow-up spectroscopic survey: i.e., on the observability
threshold. We considered the same 20 realizations as in the
previous section, but now with different limits on the apparent
magnitude of observable galaxies: mr = 24, 25, 26. The case
Petiteau, Babak, & Sesana
(d) apparent magnitude <26
20
15
index
10
5
0.3 -0.3 -0.2 -0.1 0 0.1 0.2 0.3
(f) improved WL + merger
20
15
index
10
-0.3 -0.2 -0.1 0 0.1 0.2 0.3
w, δw
mr = 26 is given in panel (d) of Figures 6 and 7. The results are
comparable to the fiducial case. They show a small improvement
in sigma and slightly larger bias for the combined distribution.
We also notice that four out of five “bad” cases remain bad.
We should say few words about the number of galaxies used
here. As mentioned above, the typical number of galaxies for
the fiducial case (mr = 24) is less than few thousand for events
at z < 1 and less than few tens of thousands for the high-redshift
event. For the improved observational limit (mr = 26), these
numbers are 2–10 times larger. The fact that our results are not
sensitive to the depth of the survey reflects the self-similarity of
the spatial distribution of galaxies in different mass ranges.
5.5. Improving the Sky Localization and the Luminosity
Distance Estimation
In our fiducial setup, the assumed source sky localization and
luminosity distance error are rather conservative. In this subsec-
tion, we consider the effect of improving such measurements.
So far, we considered only the inspiral part of the GW signal;
the inclusion of merger and ringdown will improve the localiza-
tion of the source by at least a factor of two (McWilliams et al.
2010), due to the large gain in S/N. We artificially reduced the
sky localization error coming from the inspiral by a factor of
two (factor of four in the area), assuming that this will be the
case if we take the full GW signal. We re-analyzed the same
20 realizations with this new error on the sky. Because the size
of the error box is smaller, the number of potential counter-
parts is reduced by a factor of ∼4 compared with the fiducial
case. The results are presented in panel (e) of Figure 6. We see
that the main effect of a better GW source localization is to re-
duce the number of outliers and to remove the non-Gaussian tails
in the combined probability. As it is clear from panel (e) of
Figure 7, the main gain comes from improvement of the “bad”
cases.
We now consider another estimation of the mean WL con-
tribution to the luminosity distance error, given in Wang et al.
(2002; green square-line curve on Figure 2). We take this in
combination with improved source localization on the sky com-
ing from taking into account the merger (as discussed above).
9
The Astrophysical Journal, 732:82 (11pp), 2011 May 10 Petiteau, Babak, & Sesana

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4

0.03 0.03

0.025

0.02 0.02

0.015

0.01 0.01

0.005

0 0
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
w w
Figure 8. Each row of panels shows our self-similarity check for a selected realization. In each row, the solid curve on the left panel corresponds the final posterior

k (w).
P (w). The solid curves on the right panel are the posteriors after removing one event, P
(A color version of this figure is available in the online journal.)

We consider the same 20 realizations. Results are shown in panel
(f) of both Figures 6 and 7. The improvement with respect to all
the other cases is obvious. Because the marginalized likelihood
πj coming with each galaxy is narrower due to the smaller error
in the luminosity distance, the final posterior on Pj (w) is also
narrower. The standard deviation σw is improved by more than
40% as compared with the fiducial case. The non-Gaussian tails
have almost completely disappeared, due to the removal of the
outliers (further improvement of the “bad” cases, the remain-
ing bad case will be treated in the Section 5.6; see also the top
panel of Figure 8). With this model of the mean WL contribu-
tion and assuming the full GW signal, the estimation of w is
improved by a factor of ∼8 as compared to the initial uniform
prior.
5.6. Consistency Check
As we mentioned above, some nearby GW event could seri-
ously bias the final posterior. We also mentioned that the odds
for the host to be in a low-density region of the universe are not
small. The posterior probability P (w) reflects the distribution
of the mass defined by the astrophysical prior pj (θ, φ, z). A
nearby GW event hosted in the low-density environment could
seriously damage the final result. An example is given in the top
left panel of Figure 8. In order to eliminate or at least test such
unfortunate cases, we performed a self-consistency test on our
results. Basically, we remove one GW event from the analysis
and see if the resulting posterior P
k (w) distributions are con- figuration equivalent to our improved case, i.e., better WL
sistent. We defined the posterior of all the events minus one as
 panel (f) of Figure 6). In both cases, the difference between our
p0 (w) j
=k Pj (s|w)
P
k (w) =   . (12)
p0 (w) j
=k Pj (s|w)dw counterparts detected) is only about a factor two.
k (w) gives similar results for all k, then we can be confident
If P
that the result is not biased by one particular unfortunate event,
and this increases our trust in the final posterior distribution.
If, conversely, all P
k (w) but one are consistent, then we say straining cosmological parameters using LISA observations of
that this one event is not in line with the remaining events and
should be abandoned. In the top panels of Figure 8, we see that
removing one event at low redshift changes the final probability
completely; the solid (red) line in the right panel is the new
posterior distribution, consistent with the true value w = 0.
However, there are still few cases where the self-consistency
test is not conclusive, and one of them is shown in the lower
panels of Figure 8. In this case, removing one “bad” nearby
event produces the red curve centered at w = 0, but removing
another (“good”) event results in the green curve, which are
mutually not consistent at all. Since in real life we will not
know which event is “good” and which one is “bad,” we will
not be able to make a clear definite statement, and our answer
will be bi-modal with a probability attached to each mode.
5.7. Comparison with the Optimal Case: Detection of
Electromagnetic Counterparts
For comparison, we have also considered the best possible
case, in which the redshifts of the GW source hosts are de-
termined unambiguously through the identification of a dis-
tinctive electromagnetic counterpart. In this case, the redshift
of each GW event is known exactly (within negligible mea-
surement errors). Therefore, the error on w comes only from
the error on luminosity distance (GW error measurement plus
WL). Considering 20 realizations with a configuration equiva-
lent to the fiducial case (Section 5.1), the global posterior dis-
tribution is a Gaussian centered at w0 = 0 with σw = 0.021
(for comparison, see panel (a) of Figure 6). With a con-
(Section 5.5), we obtain σw = 0.012 (for comparison, see to
statistical method and the best possible case (all electromagnetic
6. SUMMARY
In this paper, we presented a statistical method for con-
spinning MBH binaries and redshift surveys of galaxies. Our ap-
proach does not require any direct electromagnetic counterpart;

10
The Astrophysical Journal, 732:82 (11pp), 2011 May 10
instead, the consistency between a few dozen GW events im-
poses constraints on the redshift–luminosity distance relation-
ship. This, in turn, allows us to estimate cosmological param-
eters. This method strongly relies on the non-uniformity (i.e.,
clustering) of the galaxy distribution within the uncertainty error
box set by LISA observations, WL and priors on the cosmolog-
ical parameters.
For this first exploratory study, we fixed all the cosmological
parameters but one, w, describing the effective equation of state
for the dark energy. We used the Millennium simulation to model
the universe at different redshifts. We used a particular (VB)
hierarchical MBH formation model to mimic the MBH binary
population observed by LISA. Using this setup, we considered
between 20 and 100 realizations of the observed LISA binary
population. We tried two different models for estimating the
error in luminosity distance due to WL, we also looked at the
effect of including merger and ringdown via improvement of
the sky localization. We checked the robustness of our final result
against different depth of future spectroscopic galaxy surveys.
Our fiducial case, based on conservative assumptions, shows
that we are able to constrain w to a 8% level (2σ ), i.e., we
improve its estimate by a factor of ∼4 as compared to the
current 95% confidence interval obtained by cross-correlating
the seven year WMAP data analysis with priors coming from
H0 measurements and barionic acoustic oscillations (Komatsu
et al. 2011). Such new measurement would be at the same
level (25% better on average) than current constraints based on
seven year WMAP data plus SNe Ia observations. The optimistic
case (smaller WL disturbance and full GW waveform) allows
us a further improvement by another factor of two, providing
a factor of ∼2.5 tighter constraint than current estimates
including SNe data. Our results are most sensitive to the WL
error (witnessing once more how critical is the issue of WL
mitigation for cosmological parameter estimation through GW
observations) and are almost independent on the depth of the
redshift survey (provided we have a reasonable number of
redshift measurements per error box).
In the majority of the realizations the most information
comes from few events at low redshift, and high-redshift events
do help in case of multimodal structures in the posterior
distribution. We suggested a self-consistency check based on
the similarity of the posterior distribution from each GW event.
This increases our confidence in the final result and allows a
reduction of the risk of incurring unfortunate outlier realizations
for which we cannot place useful constraints on w. We also
compared our statistical method to the optimal situation in which
electromagnetic counterparts to the GW sources are identified,
finding an improvement of a factor of two in the latter case. In
absence of distinctive electromagnetic counterparts, statistical
methods like the one presented here can still efficiently constrain
cosmological parameters.
Although the main result of the present paper is encouraging,
it was obtained assuming a fixed cosmological model with one
free parameter only: the w parameter describing the dark energy
equation of state. Even though we will likely have a good
knowledge of most of the other cosmological parameters by
the time LISA will fly, it is worth considering models with more
degrees of freedom. In following studies, we intend to consider
a more realistic situation by releasing other cosmological
11
Petiteau, Babak, & Sesana
parameters, testing LISA capabilities of setting constraints on
a multi-parameter model.
The work of A.P. and S.B. was supported in parts by
DFG grant SFB/TR 7 Gravitational Wave Astronomy and by
DLR (Deutsches Zentrum fur Luft-und Raumfahrt). The Monte
Carlo simulations were performed on the Morgane cluster at
AEI-Golm and on the Atlas cluster at AEI-Hannover. The
authors thank Jonathan Gair and Toshifumi Futamase for useful
discussions.
REFERENCES
Arun, K. G., Iyer, B. R., Sathyaprakash, B. S., Sinha, S., & Van Den Broeck, C.
2007, Phys. Rev., D76, 104016
Arun, K. G., Mishra, C. K., Van Den Broeck, C., Iyer, B. R., Sathyaprakash,
B. S., & Sinha, S. 2009a, Class. Quantum Gravity, 26, 094021
Arun, K. G., et al. 2009b, Class. Quantum Grav., 26, 094027
Babak, S., et al. 2010, Class. Quantum Gravity, 27, 084009
Begelman, M. C., Volonteri, M., & Rees, M. J. 2006, MNRAS, 370, 289
Bertone, S., De Lucia, G., & Thomas, P. A. 2007, MNRAS, 379, 1143
Bielby, R., et al. 2010, arXiv:1005.3028
Bower, R. G., Benson, A. J., Malbon, R., Helly, J. C., Frenk, C. S., Baugh,
C. M., Cole, S., & Lacey, C. G. 2006, MNRAS, 370, 645
Cornish, N. J., & Porter, E. K. 2007, Class. Quantum Grav., 24, 5729
Danzmann, K., & the LISA Study Team 1997, Class. Quantum Gravity, 14,
1399
Davis, M., et al. 2003, Proc. SPIE, 4834, 161
De Lucia, G., & Blaizot, J. 2007, MNRAS, 375, 2
Faber, S. M., et al. 2003, Proc. SPIE, 4841, 1657
Gair, J. R., Sesana, A., Berti, E., & Volonteri, M. 2010, arXiv:1009.6172
Gültekin, K., et al. 2009, ApJ, 698, 198
Haehnelt, M. G., & Rees, M. J. 1993, MNRAS, 263, 168
Hilbert, S., Gair, J. R., & King, L. J. 2011, MNRAS, 412, 1023
Holz, D. E., & Hughes, S. A. 2005, ApJ, 629, 15
Komatsu, E., et al. 2011, ApJS, 192, 18
Lang, R. N., & Hughes, S. A. 2006, Phys. Rev. D, 74, 122001
Lang, R. N., & Hughes, S. A. 2009, Class. Quantum Grav., 26, 094035
Lang, R. N., Hughes, S. A., & Cornish, N. J. 2011, arXiv:1101.3591
Le Fèvre, O., et al. 2003, Proc. SPIE, 4841, 1670
Le Fèvre, O., et al. 2005, A&A, 439, 845
Lilly, S. J., et al. 2009, ApJS, 184, 218
MacLeod, C. L., & Hogan, C. J. 2008, Phys. Rev. D, 77, 043512
Madau, P., & Rees, M. J. 2001, ApJ, 551, L27
Magorrian, J., et al. 1998, AJ, 115, 2285
Malbon, R. K., Baugh, C. M., Frenk, C. S., & Lacey, C. G. 2007, MNRAS, 382,
1394
McWilliams, S. T., Thorpe, J. I., Baker, J. G., & Kelly, B. J. 2010, Phys. Rev. D,
81, 064014
Oke, J. B., & Sandage, A. 1968, ApJ, 154, 21
Petiteau, A., Shang, Y., Babak, S., & Feroz, F. 2010, Phys. Rev. D, 81, 104016
Plowman, J. E., Hellings, R. W., & Tsuruta, S. 2010, arXiv:1009.0765
Riess, A. G., et al. 1998, AJ, 116, 1009
Salvaterra, R., Haardt, F., & Volonteri, M. 2007, MNRAS, 374, 761
Schlegel, D. J., et al. 2009, arXiv:0904.0468
Schnittman, J. D. 2010, arXiv:1010.3250
Schutz, B. F. 1986, Nature, 323, 310
Sesana, A., Gair, J. R., Berti, E., & Volonteri, M. 2010, Phys. Rev. D, 83, 4036
Shapiro, C., Bacon, D. J., Hendry, M., & Hoyle, B. 2010, MNRAS, 404, 858
Springel, V., et al. 2005, Nature, 435, 629
Trias, M., & Sintes, A. M. 2008, Phys. Rev., D77, 024030
Vallisneri, M. 2008, Phys. Rev. D, 77, 042001
Van Den Broeck, C., Trias, M., Sathyaprakash, B. S., & Sintes, A. M. 2010,
Phys. Rev., D81, 124031
Volonteri, M., & Begelman, M. C. 2010, MNRAS, 409, 1022
Volonteri, M., Haardt, F., & Madau, P. 2003, ApJ, 582, 559
Wang, Y., Holz, D. E., & Munshi, D. 2002, ApJ, 572, L15
White, S. D. M., & Rees, M. J. 1978, MNRAS, 183, 341
Zombeck, M. V. 1990, Science, 249, 1314